Hugging Face Releases Carbon-X, a 14B Open DNA Language Model that Brings DNA and Natural Language Together for the First Time

Tokenize</button><button class=”btn gh” id=”cx-rst”>Reset</button></div>
<p class=”note”>5,515 DNA tokens added: 4,096 six-mers, 1,364 tail tokens and 3 markers (&lt;dna&gt;, &lt;/dna&gt;, &lt;oov&gt;). Context: 40,960 tokens.</p>
</div>

<div class=”pnl” data-i=”1″>
<p class=”lead”>Built from <b>Qwen3-14B</b> in 4 stages. Click a stage.</p>
<div class=”flow” id=”cx-flow”>
<div class=”st on” data-s=”0″><div class=”n”>01</div><div class=”t1″>Warmup</div><div class=”t2″>~10B tokens</div></div>
<div class=”st” data-s=”1″><div class=”n”>02</div><div class=”t1″>Mid-training</div><div class=”t2″>302B tokens</div></div>
<div class=”st” data-s=”2″><div class=”n”>03</div><div class=”t1″>SFT</div><div class=”t2″>3 epochs</div></div>
<div class=”st” data-s=”3″><div class=”n”>04</div><div class=”t1″>RL</div><div class=”t2″>Async GRPO</div></div>
</div>
<div class=”box det” id=”cx-det”></div>
</div>

<div class=”pnl” data-i=”2″>
<p class=”lead”>Carbon-X with thinking off vs rivals at their best reasoning effort.</p>
<div class=”bsel” id=”cx-bsel”></div>
<div id=”cx-bars”></div>
<p class=”note” id=”cx-bnote”></p>
</div>

<div class=”pnl” data-i=”3″>
<p class=”lead”>Priced at measured throughput on an <b>H200 at $5/hour</b>.</p>
<div class=”grid2″>
<div class=”box”><div class=”big” id=”cx-c1″>$0</div><div class=”cap”>Estimated cost to score 2.4M ClinVar variants of uncertain significance with Carbon-X.</div></div>
<div class=”box”><div class=”big” style=”color:#a7adbb!important”>$32K to $4M</div><div class=”cap”>Same job on the frontier models Hugging Face evaluated, by model and effort.</div></div>
<div class=”box”><div class=”big” id=”cx-c2″>0x</div><div class=”cap”>Lower inference cost per problem than GPT-6.1 Sol (model card: about 1/1,800).</div></div>
<div class=”box”><div class=”big” id=”cx-c3″>+0.0</div><div class=”cap”>Points gained by Kimi K3 when it consults Carbon-X as a sub-agent (49.4% to 60.5%).</div></div>
</div>
</div>

<div class=”pnl” data-i=”4″>
<p class=”lead”>Not clinically validated. Tap a card for details.</p>
<div class=”lim”>
<div class=”box”><div class=”h”>Splice sites <em>29.8%</em></div><div class=”b”>Weakest GenomeQA task. Histone marks are also low at 34.6%.</div></div>
<div class=”box”><div class=”h”>Exact operations <em>!</em></div><div class=”b”>It can miscount bases or confuse coordinates in long sequences. Use code for counting and edits.</div></div>
<div class=”box”><div class=”h”>Hallucination <em>!</em></div><div class=”b”>It may invent a sequence’s identity without context. Give gene, strand and annotations.</div></div>
<div class=”box”><div class=”h”>Long generation <em>!</em></div><div class=”b”>Quality degrades on long DNA generation. The model card says it was not trained to design new sequences.</div></div>
</div>
</div>

<div class=”ft”><span>Sources: HuggingFaceBio/Carbon-X model card and release blog</span><span class=”mtp”>© Marktechpost</span></div>
</div>
<script>
(function(){
var R=document.getElementById(‘mtp-cx’);
function post(){try{parent.postMessage({mtpcx:1,h:R.offsetHeight+40},’*’)}catch(e){}}
var tabs=R.querySelectorAll(‘.tab’),pn=R.querySelectorAll(‘.pnl’);
tabs.forEach(function(t){t.onclick=function(){tabs.forEach(function(x){x.classList.remove(‘on’)});pn.forEach(function(x){x.classList.remove(‘on’)});t.classList.add(‘on’);var p=R.querySelector(‘.pnl[data-i=”‘+t.dataset.p+'”]’);p.classList.add(‘on’);if(t.dataset.p===’2′)drawBars(cur);if(t.dataset.p===’3′)countUp();setTimeout(post,60)}});
// tokenizer
var TOK=[[‘Describe’,’t’],[‘this’,’t’],[‘sequence’,’t’],[‘:’,’t’],[‘<dna>’,’m’],[‘ACGTAC’,’d’],[‘GATTCA’,’d’],[‘TATAAA’,’d’],[‘</dna>’,’m’]];
var box=document.getElementById(‘cx-toks’),timer=[];
function build(){box.innerHTML=”;TOK.forEach(function(k){var s=document.createElement(‘span’);s.className=’tk ‘+k[1];s.textContent=k[0];box.appendChild(s)})}
build();
document.getElementById(‘cx-tok’).onclick=function(){timer.forEach(clearTimeout);build();var el=box.children;for(var i=0;i<el.length;i++)(function(i){timer.push(setTimeout(function(){el[i].classList.add(‘show’)},120*i+40))})(i)};
document.getElementById(‘cx-rst’).onclick=function(){timer.forEach(clearTimeout);build()};
// training
var ST=[
{h:’Embedding warmup’,p:’New DNA tokens get random embeddings. About 10B tokens train only those embeddings, with the transformer body frozen.’,mix:null},
{h:’Mid-training: 302B tokens’,p:’Sequence length 8,192, cosine LR 5e-5 to 5e-6. A lower LR, a 14B model and a long run kept English while learning DNA.’,mix:[[55,’#ffb84d’,’Raw DNA 55%’],[20,’#7cc4ff’,’English 20%’],[15,’#FF9D00′,’Biology text 15%’],[10,’#FFD21E’,’Instruction 10%’]]},
{h:’Supervised fine-tuning’,p:’3 epochs over 20 dataset subsets. Variant-effect reasoning is about 30% of the mix. Answers were checked by deterministic verifiers built with domain experts.’,mix:null},
{h:’Reinforcement learning’,p:’Asynchronous GRPO with code-checked rewards, 2 rounds, 16 samples per prompt. RL alone on Qwen3-14B never beat the 50% chance baseline on FrontierVEP.’,mix:null}];
var det=document.getElementById(‘cx-det’),sts=R.querySelectorAll(‘.st’);
function showSt(i){sts.forEach(function(s){s.classList.toggle(‘on’,s.dataset.s==i)});var s=ST[i],h='<div style=”font-weight:700;font-size:15px;color:#FFD21E”>’+s.h+'</div><div class=”cap” style=”font-size:13px”>’+s.p+'</div>’;
if(s.mix){h+='<div class=”mix”>’+s.mix.map(function(m){return ‘<i style=”background:’+m[1]+'” data-w=”‘+m[0]+'”></i>’}).join(”)+'</div><div class=”leg”>’+s.mix.map(function(m){return ‘<span style=”–c:’+m[1]+'”>’+m[2]+'</span>’}).join(”)+'</div>’}
det.innerHTML=h;if(s.mix)setTimeout(function(){det.querySelectorAll(‘.mix i’).forEach(function(x){x.style.width=x.dataset.w+’%’})},40);setTimeout(post,60)}
sts.forEach(function(s){s.onclick=function(){showSt(+s.dataset.s)}});showSt(0);
// benchmarks
var M=[‘Carbon-X’,’GPT-6.1 Sol (Max)’,’Opus 5.5 (Max)’,’Gemini 4 Argon (Low)’,’Kimi K3 (Low)’,’DeepSeek V4.1 Flash (Low)’,’GLM-5.3 (Low)’];
var B={‘Average’:[66.7,67.5,60.9,49.8,50.4,47.1,39.8],’BioReason Coding’:[74.9,71.0,70.3,6.5,53.9,51.4,52.1],’BioReason Non-SNV’:[85.1,91.7,94.8,90.3,65.3,55.9,60.1],’GenomeQA’:[63.6,67.2,62.2,63.1,52.0,49.7,44.6],’NextBase’:[39.4,32.9,1.1,31.2,30.6,27.2,9.7],’FrontierVEP’:[70.4,74.9,76.2,58.0,50.1,51.3,32.7]};
var N={‘Average’:’Carbon-X trails GPT-6.1 Sol by 0.8 points and sits 27 points above Qwen3-14B.’,’BioReason Coding’:’Coding-variant effects, 1,018 variants. Carbon-X leads every model tested.’,’BioReason Non-SNV’:’Insertions, deletions and other non-SNV changes, 873 variants. Opus 5.5 leads.’,’GenomeQA’:’2,600 multiple-choice questions over 6 tasks. GPT-6.1 Sol leads.’,’NextBase’:’Next-base prediction, 600 sequences. Carbon-X leads; Opus 5.5 collapses as refusals rise.’,’FrontierVEP’:’Single-base variants with up to 8,192-base context, 1,000 variants. Opus 5.5 leads.’};
var cur=’Average’,bs=document.getElementById(‘cx-bsel’),bars=document.getElementById(‘cx-bars’);
Object.keys(B).forEach(function(k){var b=document.createElement(‘button’);b.textContent=k;if(k===cur)b.className=’on’;b.onclick=function(){bs.querySelectorAll(‘button’).forEach(function(x){x.className=”});b.className=’on’;cur=k;drawBars(k)};bs.appendChild(b)});
function drawBars(k){var v=B[k];bars.innerHTML=M.map(function(m,i){var me=i===0?’ me’:”;return ‘<div class=”bar”><div class=”nm’+me+'” title=”‘+m+'”>’+m+'</div><div class=”tr”><div class=”fl’+me+'” data-w=”‘+v[i]+'”></div></div><div class=”v”>’+v[i].toFixed(1)+'</div></div>’}).join(”);
setTimeout(function(){bars.querySelectorAll(‘.fl’).forEach(function(f){f.style.width=f.dataset.w+’%’})},40);document.getElementById(‘cx-bnote’).textContent=N[k];setTimeout(post,60)}
drawBars(cur);
// cost
function anim(el,to,fmt){var t0=null;function f(ts){if(!t0)t0=ts;var p=Math.min(1,(ts-t0)/1000),e=1-Math.pow(1-p,3);el.textContent=fmt(to*e);if(p<1)requestAnimationFrame(f)}requestAnimationFrame(f)}
function countUp(){anim(document.getElementById(‘cx-c1’),300,function(x){return ‘$’+Math.round(x)});anim(document.getElementById(‘cx-c2’),1800,function(x){return ‘~’+Math.round(x).toLocaleString(‘en-US’)+’x’});anim(document.getElementById(‘cx-c3’),11.1,function(x){return ‘+’+x.toFixed(1)})}
// limits
R.querySelectorAll(‘.lim .box’).forEach(function(b){b.onclick=function(){b.classList.toggle(‘open’);setTimeout(post,380)}});
window.addEventListener(‘load’,post);window.addEventListener(‘resize’,post);setTimeout(post,300);
})();
</script>
</body></html>
” style=”width:100%;height:600px;border:0;display:block;background:transparent;” loading=”lazy”>